"On-demand GPU rental capacity is sold out across all GPU types…"
The market for on-demand Graphics Processing Unit (GPU) rentals is entirely depleted, irrespective of the specific GPU model. Those already secured with on-demand instances show no inclination to release their capacity, even with recent price increases. This suggests that, in the on-demand segment, how much is being used, rather than price, acts as the most telling sign of demand.
The distinction between on-demand pricing and longer-term rental agreements is significant. Short-term rentals typically encompass 'on-demand', 'spot', and contracts lasting less than three months. In contrast, mid-term arrangements span from three months up to three years or more. The new 'SemiAnalysis H100 1-year rental contract price index' is designed to specifically track the contract market, which is where the bulk of GPU rental transactions occur.
Frequently Asked Questions
Q: Why are on-demand GPU rentals completely sold out?
The demand for GPUs is very high, and all available on-demand rental capacity is currently used. Companies are keeping their rented GPUs even though prices have gone up.
Q: Who is affected by the GPU rental scarcity?
Businesses and researchers needing GPUs for short periods, like for AI projects or testing, are affected. They cannot get the computing power they need when they need it.
Q: What is the difference between on-demand and longer-term GPU rentals?
On-demand rentals are for short periods (less than three months). Longer-term rentals are for three months to three years or more. Most GPU rentals happen in longer contracts.
Q: How does this GPU shortage impact AI and machine learning?
AI and machine learning projects need a lot of computing power from GPUs. When GPUs are not available, these projects can be delayed or stopped, slowing down innovation.
Q: When will more GPU rental capacity be available?
The article does not say exactly when new capacity will be available. It notes that the market for longer-term contracts is tracked separately.
